SCEC Award Number 12022 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title An Event-Based Earthquake Predictability Experiment
Investigator(s)
Name Organization
Allan Rubin Princeton University
Other Participants Maximilian J. Werner, postdoctoral researcher at Princeton University will be very involved in proposal and will be funded.
collaborations will take place w/Dr. David D. Jackson and Dr. Yan Y. Kagan both at UCLA
SCEC Priorities 2b, 2d SCEC Groups CSEP, EFP, WGCEP
Report Due Date 03/15/2013 Date Report Submitted N/A
Project Abstract
 To investigate the predictive skills of earthquake forecasting models, we developed a suite of event-based and short-term forecasting models for California. The event-based forecasts are specified by the conditional intensity function of the ETAS model and evaluated using their continuous and exact likelihood functions, circumventing the need for discrete spatio-temporal bins and a Poisson-approximated likelihood that are currently used in CSEP experiments. We evaluated the influence of eight popular spatial triggering kernels on the probability gain and found that power-law kernels with a scale parameter that grows with mainshock rupture length work best. Lowering the learning catalog threshold to m2+ also improves forecasts of target earthquakes m3.95+, providing further evidence that small quakes improve the predictive skill of clustering models. To understand the influence of forecast horizon on probability gain, we also developed sub-24 hour (discrete) forecasts based on the ETAS model [Werner et al., 2011] and two new models K2 and K3 [Helmstetter and Werner, 2013, in review], which are based on adaptive kernel smoothing of seismicity in time, space and magnitude. 1-hour forecasts reach gains of 200 over time-independent forecasts -- significantly larger than the gains of 55 of 24-hour forecasts. Since installation of ETAS and K3 within CSEP in September 2012, 30 earthquakes m3.95+ have occurred. Although it is too early to judge the practical significance of these results, the models thus far perform better than any other 1-day model, including the STEP model by Gerstenberger et al. (2005) and the critical-branching model by Kagan and Jackson (2010).
Intellectual Merit Quantifying the predictability of earthquake occurrences in California is a major research objective of the CSEP initiative within SCEC. We have shown quantitatively that the probability gains of Omori-Utsu clustering models increase dramatically as the updating frequency of short-term forecasts is increased. To probe nuanced hypotheses of earthquake triggering that require high updating frequencies (such as dynamic stress triggering), simple reference models need to be available to establish the predictive skills of such hypotheses. Our 30-minute ETAS forecasts, for example, are ready for installation within CSEP once this forecast group becomes available (currently in development). Such short-term forecasts can also compete against Operational Earthquake Forecast (OEF) models, such as the prototyped UCERF3 spatio-temporal clustering model, to help quantify and validate the performance of OEF model candidates.
Broader Impacts Government agencies in several countries are developing capabilities for Operational Earthquake Forecasting (OEF), that is, to provide authoritative real-time information about time-varying seismic hazard to the public. Candidate models for OEF purposes need to be continuously validated and tested against other available models to ensure that the best-available science is used to inform the public about the time-dependence of earthquake hazards. This project has contributed to our understanding of the short-term predictability of earthquakes and increased CSEP capabilities to support OEF efforts by providing well-calibrated reference models.
Exemplary Figure Figure 3: Information gains per earthquake of forecasts by the ETAS model for Californian m3.95+ earthquakes during 1992 to 2012 for various updating frequencies (forecasting time horizons). Forecasts substantially improve when forecasts are updated more frequently. Anisotropic spatial aftershock kernels (circles) also improve forecasts over isotropic aftershock kernels (squares), as does lowering the magnitude threshold of the learning catalog (colors). [Werner and Helmstetter, 2013, in preparation]